复杂环境下红外强度-偏振图像融合中特征信息分离方法研究

The effective separation of image feature information is the key to realize the effective fusion of infrared intensity and polarization images under the condition of unknown scene. However, due to the influence of environment and scene in the areas with different complex environment , the difference of feature information among different images is very significant, resulting in the following two challenges for the existing image fusion methods: one is the lack of robust and effective image details extraction method, and the other is the lack of effective separation method for image differential feature information. As a result, aiming to the effective fusion for infrared intensity and polarization images, and encompassing these two constrained problems, this study will propose the following researches: (1) based on empirical mode decomposition with gradient descent and adaptive interpolation function, obtaining the details feature information from images; (2) based on sparse representation, separating the different features information between images. In the framework of mixed theory, the study innovatively designs an effective and robust extraction method for detail feature information and a effective separation method for differential feature information. It can further enrich image fusion models based on empirical mode decomposition and sparse representation for infrared identification.image fusion theory, the study innovatively designs an effective and robust detail information extraction method and a effective different information separation method. It can further enrich image fusion models based on empirical mode decomposition and sparse representation for infrared identification.

图像特征信息有效分离是实现未知场景情况下红外强度与偏振图像有效融合的关键。然而，在不同的复杂环境下，受环境和场景条件的影响，图像之间的特征信息差异性极为显著，导致现有图像融合方法面临两个问题：一是缺乏鲁棒且有效的图像细节特征信息分离方法，二是缺乏有效的图像差异特征信息分离方法。因此，本项目以红外强度与偏振图像有效融合为目标，围绕这两个约束问题展开如下研究：（1）基于梯度下降法和自适应插值函数联合优化经验模态分解，获取图像的细节特征信息；（2）基于稀疏表示的图像特征表征方法研究，分离图像之间的差异特征信息。创新点在于：在混合理论框架下设计了一种鲁棒且有效的红外强度与偏振图像细节特征信息分离方法和一种有效的图像差异特征信息分离方法，进一步丰富了基于稀疏表示和经验模态分解开展红外目标识别的图像融合模型。

本项目以红外强度与偏振图像信息有效融合为目标，围绕图像细节特征信息有效抽取和差异特征信息有效分离开展研究。从算法设计原理和应用两方面出发，探讨了双变量经验模态分解解析多源图像细节特征信息的能力，分析了使用稀疏表示表征多源图像之间差异特征信息的可行性，研究了卷积神经网络抽取图像多尺度特征的问题，提出基于双变量经验模态分解和稀疏表示的红外偏振与强度图像融合方法和基于IY型网络结构深度学习的通用性多模态图像融合方法，为图像融合模型研究提供新方向和新思路。同时，基于图像特征信息分析相关成果，研究了基于图像特征信息对沥青路面纹理进行高精度三维重建的方法，为准确分析沥青路面抗滑性能提供了一定的技术支撑。.依托本项目，培养研究生3人，协助培养博士研究生1人、硕士研究生4人，发表SCI论文3篇、中文核心论文2篇，申请中国专利3项。